Interview, Fireside Chat
Michael Kearns: Differential Privacy
- Differential privacy is not expected to prevent all negative outcomes for individuals resulting from data analysis, but rather to ensure such outcomes would remain consistent regardless of whether a specific individual's data were included or excluded.
- The field is projected to demonstrate that privacy mechanisms are not overly restrictive, enabling the execution of most computable tasks and a vast array of statistical and machine learning methods with privacy guarantees.
- Nearly all current statistical and machine learning techniques, including backpropagation in neural networks, CART decision trees, support vector machines, and boosting, are anticipated to be adapted through noise-injection modifications to provide differential privacy.
- Classic statistical procedures, such as hypothesis testing, are expected to become performable in a differentially private manner through the strategic application of noise in various computational stages.
- A scalable path is foreseen for the scientific community to extend privacy guarantees to existing methods heavily reliant on data analysis and modeling.
- Researchers anticipate a future where individuals can continue to benefit from the data science era while receiving robust privacy protections.